An integrated implementation of probabilistic graphical models

• Ano de defesa: 2020
• Autor(a) principal: Ferreira, Renato Cordeiro
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Aprendizagem de máquina; Arquitetura de software; Design patterns; Machine learning; Modelos Probabilísticos Gráficos; Padrão secretário; Padrões de projeto; Probabilistic Graphical Models; Secretary pattern; Software architecture
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/45/45134/tde-19062025-195033/

Probabilistic Graphical Models (PGMs) are a class of machine learning models used for sequence labeling and alignment. They are widely applied in many research fields, such as natural language processing, speech recognition, computer vision and bioinformatics. Firstly, this project provides a review about PGMs. It summarizes the relationship between 11 models studied and introduces a standardized representation of their algorithms including a comparison of their time and space algorithmic complexity. Thereafter, this project presents two new models that integrate the implementation of PGMs: the Generalized Multi-Sequence Hidden Markov Model (GMHMM), and the Generalized Covariance Model (GCM). It introduces their algorithms and shows how to configure the graph structure of the models to achieve the same algorithmic efficiency as the specialized models they extend. Finally, this project resulted in the extension and major refactoring of the Toolkit of Probabilistic Models of Sequences (ToPS) framework. It shows how the refactoring of the system led to a novel architecture to implement PGMs and the discovery of a new object-oriented design pattern called the Secretary pattern.